Estimating Body Pose from A Single Image

The Problem: The problem is to estimate the body pose of a person from a single picture taken from an uncallibrated camera. Motivation: The specific problem is motivated by numerous applications that fall in the general categories of human computer interaction, video surveillance, virtual reality, computer animation, and others. Another motivation is the development of general computer vision techniques for matching and recognition of articulated and flexible models. Previous Work: There are a number of systems that tackle the specific problem defined here, the latest one being [2]. They use fairly generic machine learning techniques to find a mapping from image silhouettes to body pose. Our work takes a more geometric approach by modeling the body as an articulated model. There are a number of techniques for matching articulated models to images, see [1] for a brief overview. Approach: We started with the techniques developed in [1]. Those techniques allow us to model an object as a collection of parts connected by joints. Our body model consists of rigid body parts (torso, head, limbs) and flexible revolute joints connecting the parts in a way to approximate the human skeleton. We use generic apperance models for the individual body parts. The characteristics of the joints connecting different body parts encode a prior model of “typical ” body poses. This prior model is learned from examples (motion capture data available on the web). Figure 1 shows samples from the learned probability distribution over body poses. The algorithms in [1] allows us to efficientl